Concrete made with industrial byproducts and recycled aggregate is assembled from parts that vary a great deal: a particular fly ash, a particular slag, a recycled aggregate from one demolition site, a given admixture, a given curing age. Each combination of these choices behaves like its own small population, and most combinations have been tested only a handful of times, or never. Building codes do not ask a strength model for its average error. They ask for a characteristic strength, a value that roughly 95% of specimens of that mix should exceed. Current machine-learning predictors report excellent average fit, but they say nothing reliable about the lower tail of a rare combination. In materials science they are also known to fail sharply on unfamiliar feedstocks.
This proposal treats feedstock combinations the way the fairness literature treats intersectional demographic groups. These are cells formed by crossing several attributes, and they quickly become too sparse to estimate one at a time. We would build a one-sided prediction bound, meaning a guaranteed-style lower limit on strength, for each feedstock cell. The bound would borrow calibration evidence from neighboring cells that share some of the cell's attributes, and it would report how much of the bound comes from the cell's own tests. The aim has two parts. First, rare cells should get bounds that are both non-trivial and close to the 95% target. Second, cells that are genuinely worse than their neighbors should be flagged rather than hidden by the borrowing.
Core Claim
Calibration that is pooled over the whole dataset will under-protect rare feedstock combinations. Here, calibration means using held-out residuals (prediction errors on samples the model did not train on) to set an error margin. Calibration done separately inside each cell is valid but useless for small cells. With a one-sided 5% target, a cell with fewer than about 19 calibration samples gets an infinitely loose bound. That threshold comes from the standard finite-sample conformal quantile rank.
We claim that weighting calibration residuals by how many attributes a donor cell shares with the target cell gives the best of both. The weights would be fit by empirical Bayes, meaning they are estimated from the observed between-cell variation in residuals. Rare cells would get coverage close to 95% with bounds tight enough to use, and the method would carry an explicit "own-evidence fraction" that tells an engineer when a bound rests mostly on other mixes. The new element is not conformal prediction or hierarchical shrinkage on its own. It is joining the two on a lattice of material attributes, with a target set by building codes and a built-in check against borrowing that hides real defects.
Where This Sits in Prior Work
The materials side supplies the motivation. Tree ensembles and boosted neural networks fit the 1,030-mix UCI concrete dataset very well. Wang et al. report XGBoost at $R^2=0.95$ with a 3.5% average error on cast blocks, and ISSA-BPNN-AdaBoost reaches $R^2=0.964$. Both optimize a single global accuracy number and report no lower-tail or per-subgroup behavior; we target exactly that missing quantity.
The UHPC inverse-design study is the closest materials work. It shows that a compositional Gaussian process falls from test $R^2=0.91$ to $R^2=-0.06$ on mixes with novel sand and limestone filler. It also shows that physically grounded intermediate variables, such as particle packing and water film thickness, recover $R^2=0.67$ and cut miscalibration area from 0.20 to 0.06. We reuse those variables as the base predictor's features rather than competing with them. That study offers no per-combination guarantee, and we add one.
Jayaram et al. cluster 72 fly ashes into three chemical groups. Their work shows that secondary markers such as $\text{SO}_3/\text{TiO}_2$ matter, which is part of why attribute crossings, and not only bulk chemistry, define meaningful cells. Their grouping is unsupervised and coarse, and it carries no uncertainty.
The intersectional-sparsity side supplies both constraints and methods. Paes et al. prove that per-group auditing becomes impossible once group count outpaces samples per group. They also show that models trained to minimize average loss can be worse for specific groups than a generic model. The first result constrains our evaluation: we audit coverage over families of cells grouped by size, not cell by cell. SAFT finds that frequentist and Bayesian intervals agree only above roughly 30–50 samples, which fixes the regime where borrowing must do the work. DRAF scales subgroup fairness by ignoring subgroups below a support threshold. That policy is unacceptable when a rare waste stream is about to be poured, so it serves as our "refuse to predict" reference.
The method comes from the shrinkage line. Hierarchical Beta Calibration cuts fairness-metric error on German Credit from about 30% to about 8% at 10 labels by sharing strength across groups. Structured embeddings share parameters across 83 party–state groups through hierarchical priors. Neither produces prediction bounds, and neither handles continuous physical inputs. Conformal Fairness calibrates group-specific prediction sets and reduces coverage disparity on ACSIncome from above 0.40 to below 0.05. However, it calibrates each group with that group's own data, which is exactly what breaks for cells of size 3.
Research Program
The base predictor is any point model $f$, such as XGBoost or a Gaussian process on mechanistic features. Each mix $x$ belongs to a cell $g(x)$ defined by $q$ attributes. On UCI, these attributes are binned slag content, binned fly ash content, superplasticizer presence, water-to-binder bucket, cement bucket, and age bucket, which gives a few hundred possible cells, most of them sparse. On the UHPC data, the publication source is added as an attribute.
For a test mix in cell $g$, the bound is
$$L(x) = f(x) - \hat q_g, \qquad \hat q_g = \text{weighted } (1-\alpha)\text{ quantile of } \{\, y_i - f(x_i) \,\}_{i \in \text{cal}} \text{ with weights } w_{s(i,g)}.$$
In this formula, $\alpha=0.05$ is the code-style 5% exceedance level. The term $y_i - f(x_i)$ is how far a calibration mix's measured strength fell below or above its prediction. The quantity $s(i,g)$ counts how many of the $q$ attributes calibration mix $i$ shares with cell $g$, and $w_s$ is a weight that grows with $s$. Own-cell residuals have $s=q$.
The weights $w_s$ are fit by empirical Bayes from how much residual quantiles actually vary between cells at each sharing level. Where cells sharing $s$ attributes have similar residual tails, they receive near-full weight. Where tails diverge, their weight decays toward zero.
The own-evidence fraction for cell $g$ is
$$\rho_g = \frac{n_g w_q}{\sum_s n_{g,s} w_s}.$$
Here $n_g$ is the number of calibration mixes in cell $g$ itself, and $n_{g,s}$ is the number of calibration mixes sharing exactly $s$ attributes with $g$. It is reported alongside the bound. A low $\rho_g$ tells the engineer that the cell's bound is mostly borrowed and that more destructive tests are needed before relying on it.
Exact finite-sample validity is lost once foreign residuals are mixed in. Standard weighted-conformal arguments bound the resulting coverage shortfall by the distance between the weighted donor residual distribution and the cell's true one. We would state this bound and then measure the actual shortfall empirically rather than rely on it.
The analysis has three parts. We compare per-cell coverage across cell-size bins. We measure tightness, defined as the mean MPa gap between the bound and the true characteristic strength. And we stress-test the method against cells that are genuinely deficient.
Baselines
The naive reference is the point predictor minus a single global residual quantile, that is, split conformal with no groups. It has the most calibration data, so it gives the tightest bounds. It has exact marginal coverage, but we expect it to under-cover rare and atypical cells. We will run it on identical train, calibration, and test splits and the same base model. The proposal beats it if worst size-bin coverage on rare cells (fewer than 10 calibration mixes) rises toward 95% while the overall average bound loosens by only a small, reported amount.
The strongest existing per-group method is Mondrian conformal, which calibrates each cell separately. It is the natural transplant of the Conformal Fairness idea to regression. It is the strongest competitor on validity because its coverage holds exactly within every cell. We run it on the same cells, and where a cell has too few calibration samples its bound is reported as trivial. We beat it if, at matched coverage within two points, we certify a much larger fraction of rare cells with finite bounds and those bounds are tighter.
A second strong competitor is Mondrian calibration on marginal attributes only. It uses one attribute at a time, so each group is large; this corresponds to the parent level of our lattice. It isolates whether the intersection structure is needed at all. We beat it only if it under-covers specific intersections that our method covers.
For point accuracy, the reference is XGBoost as in Wang et al., who report $R^2=0.95$ on UCI. The strongest point predictor in the file is ISSA-BPNN-AdaBoost at $R^2=0.964$, and we will match base-model accuracy to the 0.95–0.96 range so the comparison is about calibration, not fit. On novel UHPC materials, the reference is the latent-variable Gaussian process at $R^2=0.67$ from the UHPC study. No reported result was found on Lacuna for per-group or lower-bound coverage on any concrete dataset, so all coverage baselines will be run by us.
The DRAF-style "refuse below threshold" policy from DRAF is reported as the fraction of mixes it declines to bound at thresholds of 30 and 50.
The key ablation fixes uniform weights across sharing levels, $w_s$ constant for all $s$. This collapses the method to global conformal with cell labels. A second ablation removes the empirical-Bayes fit and uses hand-set weights. Together they isolate whether learned, similarity-graded borrowing is the active ingredient. All methods share the same calibration pool and the same hyperparameter budget, which is a fixed grid over the base model only.
Why not just…?
Why not just use XGBoost with ordinary split conformal? This is the real threat on average-case metrics, and we expect it to win on overall tightness. The case against it is evidence from the file that average fit hides subgroup failure. It includes the collapse to $R^2=-0.06$ on novel UHPC materials reported in the UHPC study, and the finding by Paes et al. that models trained for average performance can be worse on specific groups. The leave-cell-out experiment settles the question directly. If global conformal already covers rare cells at close to 95%, the proposal's premise fails on this data, and we would report that.
Why not just calibrate each cell separately? It is valid, but at $\alpha=0.05$ any cell with fewer than about 19 calibration mixes gets no usable bound. On UCI that describes most cells. SAFT independently places the reliable-estimation regime above about 30–50 samples. The certified-fraction metric shows the gap.
Why not just rely on the mechanistic latent variables? They improve generalization; the UHPC study reports a validation $R^2$ of 0.67 and a miscalibration area of 0.06. But that is still a point predictor with global uncertainty, not a per-combination lower bound. We therefore use those variables inside $f$ rather than as an alternative. If a latent-feature Gaussian process with global calibration already covers rare cells, the borrowing layer adds nothing, and the UHPC leave-source-out run will show that.
Why not just refuse to bound under-tested mixes? This is the DRAF policy. It is safe but defeats the purpose of circular feedstocks, which arrive in small heterogeneous lots, as with the treated recycled aggregates in the carbide slag study. The own-evidence fraction $\rho_g$ offers a graded version of refusal instead of an all-or-nothing cutoff.
Proposed Experiments
On UCI, we use three splits. The first is a random split, as a sanity check. The second is a leave-cell-out protocol: whole intersections are held out, and only $k \in \{0, 2, 5, 10\}$ of their mixes are revealed for calibration. The third is a leave-attribute-level-out split, for example holding out all high-slag, low-age mixes. On the UHPC literature data we hold out each of the four sources in turn, plus the in-house novel-sand mix. The 76-mix Chopra data from the ISSA-BPNN study serves as a small-data stress case.
The metrics are coverage of the 5% lower bound, averaged within cell-size bins (following the auditing limits of Paes et al.), and the worst size-bin coverage. We also report mean slack in MPa between the bound and the observed fifth percentile, and the fraction of test mixes with a finite bound.
The decisive check is a planted-defect test. Into one held-out cell we inject a strength penalty of 10–20% that its neighbors do not have, standing in for an untested sulfate-rich ash. Borrowing that is too aggressive would hide this defect. The method passes only if coverage on that cell degrades no worse than under per-cell calibration with the same $k$, and if $\rho_g$ is low enough to flag the cell. Two outcomes would lead us to abandon the intersectional structure: if the marginal-only calibration baseline matches our coverage on intersections, or if uniform weights match the learned weights. In either case we would report that coarse groups suffice for this material data.
A secondary transfer check on ACSIncome, regressing income on census features, tests whether the same weighting helps rare demographic intersections. This would indicate whether the contribution generalizes beyond concrete.
References
- Research on prediction of compressive strength of fly ash and slag mixed concrete based on machine learning
- Prediction of Concrete Compressive Strength Based on ISSA-BPNN-AdaBoost
- Bayesian machine learning for inverse design of ultra-high-performance concrete
- A Three-Stage Neural Model for Attribute Based Classification and Indexing of Fly Ashes
- On the Epistemic Limits of Personalized Prediction
- Size-adaptive Hypothesis Testing for Fairness
- Doubly-Regressing Approach for Subgroup Fairness
- Can I Trust My Fairness Metric? Assessing Fairness with Unlabeled Data and Bayesian Inference
- Structured Embedding Models for Grouped Data
- A Generic Framework for Conformal Fairness
- Mechanical Properties of Recycled Concrete with Carbide Slag Slurry Pre-Immersed and Carbonated Recycled Aggregate